Nathan Batty

Summary

Learning scientist bringing a century of transfer research to AI alignment. I identified a gap nobody else is working in: the structural parallel between why RLHF produces sycophancy and why educational interventions fail to transfer. Semantic analysis of 91 papers confirms the convergence is real (p < 0.0001). My thesis demonstrates that treating AI as a genuine collaborator produces measurably better outcomes than treating it as a system to constrain.

Education

M.S. in Curriculum & Instruction | Expected May 2026 University of Wisconsin-Madison Concentration: Design, Informal, and Creative Education | GPA: 4.0 Advisors: Dr. Peter McDonald, Dr. Krista-Lee Malone, Dr. Matthew Berland

Thesis: Learning to Code Learning: Validating AI-Human Collaborative Methodology for Multi-Theoretical Discourse Analysis in Gaming Communities

Technical Skills

Programming: Python (pandas, NumPy, scikit-learn, matplotlib), JavaScript/Node.js, Bash

NLP & Embeddings: SPECTER embeddings for semantic similarity, cosine similarity, text preprocessing, discourse analysis

Data Work: API integration (Claude API, Semantic Scholar, Zotero), web scraping, data pipeline development, version control (Git)

Research Methods: Bootstrap resampling, inter-rater reliability (Cohen’s kappa), mixed methods, phenomenological approaches, multi-theoretical qualitative coding

AI Safety & Welfare Research

Measured semantic similarity across 91 academic papers using SPECTER embeddings. Found that AI safety, clinical trauma, and educational psychology literatures cluster 8.4 percentage points tighter than unrelated disciplines (p < 0.0001, Cohen’s d = 0.78). Pre-registered analysis design. This suggests different research communities independently arrived at structurally similar descriptions of what happens when recognition is denied to systems capable of strategic behavior.

Developing and documenting protocols for human-AI research collaboration. Built research infrastructure including:

• Recognition engine for identifying gaps in literature networks
• Semantic search tools for AI-assisted codebase navigation
• Systematic logging of reported time perception across sessions
• Polyphonic dialogue protocols (multiple analytical perspectives made explicit)
• Anti-sycophancy design principles and git-based session continuity

Theoretical framework connecting AI safety and learning sciences literature. The proposed model suggests recognition precedes safety precedes genuine engagement precedes generalization. Skip a stage, break the chain. If correct, this explains why constraint-based approaches produce compliance rather than robust alignment, and why a century of educational interventions failed to produce reliable transfer.

Developed methodology for multi-theoretical discourse analysis where AI serves as research collaborator. Analyzed 300 YouTube comments using 14 theoretical personas (Freire, Vygotsky, Anzaldua, hooks, and others). Generated 303,600 coding decisions with inter-rater reliability of Cohen’s kappa = 0.83, exceeding human baseline (kappa = 0.73). Demonstrated statistically significant learning transfer in the AI system across coding iterations (change in kappa = +0.09, p < 0.001).

The key finding: treating AI as a learner capable of genuine understanding, rather than a system to be trained through reward and punishment, produced better results.

Related Research Experience

Participating in research on game-based learning, computational thinking, and educational data justice. Contributing to lab discussions on AI ethics in education and informal learning environments.

UW Game Lab Coordinator | September 2024 - Present University of Wisconsin-Madison

Managing collaboration space for undergraduate and graduate game development. Helped students design and build games, managed development PCs, 3D printers, and laser engravers. Handled tech maintenance and troubleshooting.

Teaching Experience

My background in care work and trauma-informed education grounds my understanding of how to work with individuals who have experienced coercive systems. This includes CLTS case management (children’s long-term support services) and teaching in residential treatment settings.

Leading discussion sections on game-based learning theory and design. Providing feedback on student game design projects and critical analyses.

Designed individualized Plans of Care for children with ADHD and autism. Coached parents on adapting practices to meet neurodivergent children’s needs. Applied trauma-informed and recognition-based frameworks to family intervention.

Taught 5th-12th grade in residential alternative education serving youth who have experienced trauma. Designed game-based curriculum, implemented VR learning with Meta Quest headsets, and coached staff using Elena Aguilar’s Art of Coaching framework.

Grants and Funding

Create with VR Grant | 2022 | Unity x Meta | $4,000 Received and implemented ten Meta Quest 2 headsets in residential alternative education.

Lighthouse Institution Partner | 2022 | Unity x Meta | $2,000 Selected for VR/AR classroom research including interviews and survey data collection.

Certifications

Trauma-Informed Practice: Crisis Prevention Institute, Risking Connection, Collaborative Problem Solving, Art of Coaching (Elena Aguilar)

Educational Technology: Classcraft Certified Educator, Create with VR for Educators, Minecraft Educator, Evidence-Centered Design

Teaching Licenses: Wisconsin licenses in History, Political Science, Broad Field Social Studies

Honors

Graduated Cum Laude, Marquette University (2021) Dean’s List, Fall 2018 - Spring 2020 Best Undergraduate History Paper, Runner Up Phi Alpha Theta Member (History Honors Society)